AMD Ryzen AI Halo Review: $3,999 Box Trails DGX Spark 3.2x in vLLM

AMD's Ryzen AI Halo costs $3,999 for 128 GB of unified memory, against $6,950 for NVIDIA's 128 GB DGX Spark, but StorageReview measured it at 222 tok/s to the Spark's 701 on gpt-oss-120b in vLLM at batch 64.

That gap, about 3.2x in this one test (gpt-oss-120b, vLLM, 256 tokens in and out; quantization not stated), is wider than the price gap of about 1.7x; both ratios are Voltgrid arithmetic from StorageReview's results and the list prices, which are three months apart and for different storage sizes. Tom's Hardware, running llama.cpp instead of vLLM, reached the same conclusion from a different direction: the Halo's token rate is acceptable, but its time-to-first-token falls far behind GB10 machines as the context grows. Both outlets published their reviews on July 6. What has changed since is the rival's price. Voltgrid's DGX Spark review found the 128 GB Spark listed at $6,950 on NVIDIA's marketplace on Oct. 31, against the $4,699 StorageReview quoted in July for the Founders Edition, a different configuration. This review sets AMD's specifications against what those two outlets measured. Voltgrid did not test the machine.

What's New in the Ryzen AI Halo Developer Box

The Halo is AMD's own-brand AI developer box. Strix Halo silicon, sold as the Ryzen AI Max+ 395, shipped in partner mini PCs and laptops before the Halo. The Halo wraps that chip in a 150 x 150 x 45.4 mm case weighing under 1.2 kg, close to the GB10 boxes' footprint. AMD announced it beside the Ryzen AI Max PRO 400 series in May 2026, took pre-orders through Micro Center from June and put it in stores on July 10, according to StorageReview.

The chip is unchanged: 16 Zen 5 cores, a Radeon 8060S GPU with 40 RDNA 3.5 compute units and an XDNA 2 neural processing unit (NPU), rated at 120 W. AMD's spec page lists 128 GB of LPDDR5x at 8000 MT/s with 256 GB/s of bandwidth, a 2 TB SSD, one 10 Gbps Ethernet port, Wi-Fi 7 and HDMI 2.1b.

The new part is the software. The Linux build ships with AMD's ROCm stack and a set of inference tools preinstalled, plus AMD's own playbooks for local AI tasks, modelled on the guides NVIDIA publishes for the Spark, Tom's Hardware reported. The same box can be ordered with Windows 11, according to AMD's spec page.

Running 120B Models Locally on 128 GB

Memory decides which model fits, and 128 GB is enough for gpt-oss-120b, the open 120-billion-parameter mixture-of-experts model both outlets ran. AMD says the pool holds models of up to 200 billion parameters, StorageReview reported. That puts a 120B-class model on a desk without a cloud bill, the same promise the Spark makes.

What the Halo removes is the setup work. Tom's Hardware described the usual Strix Halo route as digging through scattered GitHub pages, Reddit threads and AMD documentation to line up a working stack. A buyer who wants that capacity on an x86 machine, with Windows applications beside the models, gets it configured out of the box.

Bandwidth sets speed. Each generated token reads the active weights from memory, so a machine at 256 GB/s should write a little slower than one at 273 GB/s, the DGX Spark's figure, all else equal. The difference is about 7% (Voltgrid arithmetic from AMD's and NVIDIA's figures). The measured gap is much larger. That points at compute and software rather than memory, and it shows most in prefill, the stage that reads the prompt before the first token appears.

Pros and Cons

The price and memory figures come from AMD's spec page, StorageReview and NVIDIA's marketplace listing. StorageReview ran vLLM online serving on both machines with gpt-oss-120b and recorded the CPU results under Linux. It found the Spark ahead by 2x to 4x in most serving runs at higher concurrency. Tom's Hardware ran Qwen 3.6-35B-A3B, Gemma 4 12B and gpt-oss-120B in Unsloth's Q4_K_M GGUF builds through llama.cpp, after the preinstalled vLLM would not start the Qwen model. It also reported the temperatures and the fan noise, which grew louder than the Dell Pro Max GB10's under an image-generation load.

Who Should Buy the Ryzen AI Halo

The Halo fits a developer who needs 128 GB on x86 and Windows, for example to run a 120B model beside Windows-only tools, and who would otherwise build the same stack on a partner Strix Halo box. It also fits a team whose code is already on ROCm, and a buyer whose budget stops near $4,000 now that the 128 GB Spark lists at $6,950.

It fits a single user with short prompts better than a long-context one. Tom's Hardware found the token rate tolerable once output starts. The wait before it starts is the problem at long context.

It does not fit a team serving one model to many users. At batch 64 on gpt-oss-120b, Voltgrid's arithmetic from StorageReview's batch-64 results and the list prices gives the Halo about 55 tok/s per $1,000 and the Spark about 101 tok/s per $1,000 (StorageReview did not publish these ratios), even at the Spark's higher price. Long-context coding agents, where prompts keep growing, sit on the wrong side of the 8.8x prefill gap.

Ryzen AI Halo vs DGX Spark and Rivals

Against the DGX Spark, the Halo has the same 128 GB, slightly less bandwidth at 256 GB/s against 273 GB/s, and a list price $2,951 lower (Voltgrid arithmetic; the Spark listing is 4 TB, the Halo 2 TB, and the prices date from July and October). StorageReview's same-lab runs on gpt-oss-120b at batch 64 put the Spark ahead by about 3.2x with equal 256-token input and output, about 2.4x on decode-heavy runs (127 against 305 tok/s at 1k in, 8k out) and about 8.8x on prefill-heavy runs.

Against the Dell Pro Precision with GB10 (sold earlier as the Dell Pro Max with GB10), $6,593 for 128 GB and 2 TB on Dell.com on Oct. 30, the picture is similar. StorageReview measured the Dell at 670.83 tok/s on the equal-length gpt-oss-120b run at batch 64 in March, against the Halo's 222 tok/s in July. Those runs came months apart, so software versions may differ. Tom's Hardware used the Dell as its GB10 reference and found it faster than the Halo on token rate and far ahead on first-token latency.

The Framework Desktop uses the same Ryzen AI Max+ 395 with 128 GB at 256 GB/s for $3,449 as a DIY Edition, but Framework listed those configurations out of stock on Oct. 30. Against NVIDIA's RTX Spark laptop, Tom's Hardware's AgentPerf-local run of Qwen 3.6 35B A3B in llama.cpp took 13 minutes on the Halo and 8 minutes on the RTX Spark Surface Laptop Ultra (Q4_K_M with MTP; lower is better; a desktop and a laptop differ in class).

The table under this review shows each machine's sourced memory, bandwidth, price and benchmark. Read together, it says the Halo is the cheapest ready-made 128 GB box with a vendor software stack, and the slowest of the three machines in the outlets' multi-user serving runs, though the Dell figure comes from a March test and Framework's identical chip was not measured here.

What to Watch for the Ryzen AI Halo

AMD reports fiscal third-quarter 2026 results on Nov. 3, after the U.S. market closes. Any remarks on the Halo, its supply or the Ryzen AI Max PRO 400 series are the next data point. AMD's PRO 495 supports up to 192 GB of unified memory and was due from HP and Lenovo in the third quarter, which has now ended. Framework opened pre-orders for a 192 GB Desktop with the PRO 495 on Sept. 30, from $6,799, shipping in November. AMD has not given a price or date for a 192 GB Halo.

Two measurements are still missing from the record: wall power under an LLM load, and a vLLM or SGLang rerun on a current ROCm release that would show whether software has closed any of the prefill gap. Voltgrid will update this review when either appears.